DALL-E

By Satwik · February 5, 2026

DALL-E (early 2021) demonstrated text-to-image generation from a transformer trained to model a single stream of text tokens followed by image tokens. Images were discretized into a codebook of tokens by a learned autoencoder, and the transformer learned the joint sequence autoregressively. At generation time you feed a caption and the model produces the image tokens, which are decoded back to pixels. Candidates were often reranked using CLIP to pick the best match to the prompt.

The demonstrations were the point. DALL-E rendered plausible images for unusual compositions, "an armchair in the shape of an avocado," combining concepts it had not seen together. This compositional generalization suggested the model had learned reusable visual concepts rather than memorized templates.

Reading angle

DALL-E marked the arrival of controllable synthetic imagery from plain language, which is a capability with obvious dual-use character. Generated media raises provenance and authenticity concerns, and the ability to conjure arbitrary scenes on demand lowers the cost of producing misleading visuals. OpenAI's initial release was cautious and limited, reflecting awareness of these issues. For a security reader, DALL-E is best noted as the moment where synthetic content generation became fluent enough that downstream authentication, watermarking, and detection became live problems rather than hypotheticals. The technical lesson, that a generic sequence model handles pixels as readily as words, also foreshadowed the unified multimodal systems that followed.